In this paper, an adaptive approach using the least-mean-square lattice (LMSL) is proposed for the segmentation and tracking of the electro-encephalogram (EEG) signal. In the proposed approach, the time-trajectories of the reflection coefficients of the adaptive lattice predictor as well as the on-line power spectrum estimate are used as classification, segmentation and tracking parameters. The adaptive lattice predictor consists of cascaded similar first-order sections. Theses sections are independent due to the orthogonality principle linked to the least-mean square (LMS) algorithm. Therefore, on-line adding a new section makes no influence on the values of the coefficients of the preceding sections. Such property of the adaptive lattice predictor is not valid when the direct-form linear prediction filter is employed. Further advantage of the lattice predictor is the fast convergence due to independence of the successive sections, which yields no matrix computation. Results for tracking sleep spindles of computer-generated and real-world EEG data are presented to show the significant usefulness of the proposed approach. It is shown that an adaptive lattice predictors consisting of three up to four sections are satisfactory for the detection and tracking of the sleep spindles. A short-time (i.e., sliding window) implementation of Burg’s method (STBM) for computing the reflection coefficients of the lattice predictor is also suggested and examined. This method shows that the time-trajectories of the lattice coefficients, the on-line power spectrum and a proposed quantitative measure are capable of distinguishing between the EEG from normal healthy subject and the EEG from Alzheimer patient subject.